---
language:
- en
---


# Emu2-Gen

[Paper](https://arxiv.org/abs/2312.13286) | [🤗HF Demo](https://huggingface.co/spaces/BAAI/Emu2) | [Demo](https://emu.ssi.plus) | [Project Page](https://baaivision.github.io/emu2/) | [Github](https://github.com/baaivision/Emu)


## Model Weights

| Model name         | Weight                                                  |
| ------------------ | ------------------------------------------------------- |
| **Emu2** | [🤗 HF link](https://huggingface.co/BAAI/Emu2) |
| **Emu2-Chat** | [🤗 HF link](https://huggingface.co/BAAI/Emu2-Chat) |
| **Emu2-Gen** | [🤗 HF link](https://huggingface.co/BAAI/Emu2-Gen) |


## Inference (Huggingface Version)

### Emu2-Gen

```python
import cv2
from diffusers import DiffusionPipeline
import numpy as np
from PIL import Image
import requests
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# For the first time of using,
# you need to download the huggingface repo "BAAI/Emu2-GEN" to local first
path = "path to local BAAI/Emu2-GEN"

multimodal_encoder = AutoModelForCausalLM.from_pretrained(
	f"{path}/multimodal_encoder",
	trust_remote_code=True,
	torch_dtype=torch.bfloat16,
	use_safetensors=True,
	variant="bf16"
)
tokenizer = AutoTokenizer.from_pretrained(f"{path}/tokenizer")

pipe = DiffusionPipeline.from_pretrained(
	path,
	custom_pipeline="pipeline_emu2_gen",
	torch_dtype=torch.bfloat16,
	use_safetensors=True,
	variant="bf16",
	multimodal_encoder=multimodal_encoder,
	tokenizer=tokenizer,
)

# For the non-first time of using, you can init the pipeline directly
pipe = DiffusionPipeline.from_pretrained(
	path,
	custom_pipeline="pipeline_emu2_gen",
	torch_dtype=torch.bfloat16,
	use_safetensors=True,
	variant="bf16",
)

pipe.to("cuda")

# text-to-image
prompt = "impressionist painting of an astronaut in a jungle"
ret = pipe(prompt)
ret.images[0].save("astronaut.png")

# image editing
image = Image.open(requests.get('https://github.com/baaivision/Emu/Emu2/examples/dog2.jpg?raw=true',stream=True).raw).convert('RGB')
prompt = [image, "wearing a red hat on the beach."]
ret = pipe(prompt)
ret.images[0].save("dog_hat_beach.png")

# grounding generation
def draw_box(left, top, right, bottom):
	mask = np.zeros((448, 448, 3), dtype=np.uint8)
	mask = cv2.rectangle(mask, (left, top), (right, bottom), (255, 255, 255), 3)
	mask = Image.fromarray(mask)
	return mask

dog1 = Image.open(requests.get('https://github.com/baaivision/Emu/Emu2/examples/dog1.jpg?raw=true',stream=True).raw).convert('RGB')
dog2 = Image.open(requests.get('https://github.com/baaivision/Emu/Emu2/examples/dog2.jpg?raw=true',stream=True).raw).convert('RGB')
dog3 = Image.open(requests.get('https://github.com/baaivision/Emu/Emu2/examples/dog3.jpg?raw=true',stream=True).raw).convert('RGB')
dog1_mask = draw_box( 22,  14, 224, 224)
dog2_mask = draw_box(224,  10, 448, 224)
dog3_mask = draw_box(120, 264, 320, 438)

prompt = [
	"<grounding>",
	"A photo of",
	"<phrase>the first dog</phrase>"
	"<object>",
	dog1_mask,
	"</object>",
	dog1,
	"<phrase>the second dog</phrase>"
	"<object>",
	dog2_mask,
	"</object>",
	dog2,
	"<phrase>the third dog</phrase>"
	"<object>",
	dog3_mask,
	"</object>",
	dog3,
	"on the grass",
]
ret = pipe(prompt)
ret.images[0].save("three_dogs.png")
```

## Citation

If you find Emu2 useful for your research and applications, please consider starring this repository and citing:

```
@article{Emu2,
    title={Generative Multimodal Models are In-Context Learners}, 
    author={Quan Sun and Yufeng Cui and Xiaosong Zhang and Fan Zhang and Qiying Yu and Zhengxiong Luo and Yueze Wang and Yongming Rao and Jingjing Liu and Tiejun Huang and Xinlong Wang},
    publisher={arXiv preprint arXiv:2312.13286},
    year={2023},
}
```